1/* 2 * Copyright © 2014 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21 * IN THE SOFTWARE. 22 * 23 * Authors: 24 * Jason Ekstrand (jason@jlekstrand.net) 25 * 26 */ 27 28#include "nir_constant_expressions.h" 29#include <math.h> 30 31/* 32 * Implements SSA-based constant folding. 33 */ 34 35struct constant_fold_state { 36 void *mem_ctx; 37 nir_function_impl *impl; 38 bool progress; 39}; 40 41static bool 42constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx) 43{ 44 nir_const_value src[NIR_MAX_VEC_COMPONENTS][NIR_MAX_VEC_COMPONENTS]; 45 46 if (!instr->dest.dest.is_ssa) 47 return false; 48 49 /* In the case that any outputs/inputs have unsized types, then we need to 50 * guess the bit-size. In this case, the validator ensures that all 51 * bit-sizes match so we can just take the bit-size from first 52 * output/input with an unsized type. If all the outputs/inputs are sized 53 * then we don't need to guess the bit-size at all because the code we 54 * generate for constant opcodes in this case already knows the sizes of 55 * the types involved and does not need the provided bit-size for anything 56 * (although it still requires to receive a valid bit-size). 57 */ 58 unsigned bit_size = 0; 59 if (!nir_alu_type_get_type_size(nir_op_infos[instr->op].output_type)) 60 bit_size = instr->dest.dest.ssa.bit_size; 61 62 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { 63 if (!instr->src[i].src.is_ssa) 64 return false; 65 66 if (bit_size == 0 && 67 !nir_alu_type_get_type_size(nir_op_infos[instr->op].input_types[i])) 68 bit_size = instr->src[i].src.ssa->bit_size; 69 70 nir_instr *src_instr = instr->src[i].src.ssa->parent_instr; 71 72 if (src_instr->type != nir_instr_type_load_const) 73 return false; 74 nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr); 75 76 for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i); 77 j++) { 78 src[i][j] = load_const->value[instr->src[i].swizzle[j]]; 79 } 80 81 /* We shouldn't have any source modifiers in the optimization loop. */ 82 assert(!instr->src[i].abs && !instr->src[i].negate); 83 } 84 85 if (bit_size == 0) 86 bit_size = 32; 87 88 /* We shouldn't have any saturate modifiers in the optimization loop. */ 89 assert(!instr->dest.saturate); 90 91 nir_const_value dest[NIR_MAX_VEC_COMPONENTS]; 92 nir_const_value *srcs[NIR_MAX_VEC_COMPONENTS]; 93 memset(dest, 0, sizeof(dest)); 94 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; ++i) 95 srcs[i] = src[i]; 96 nir_eval_const_opcode(instr->op, dest, instr->dest.dest.ssa.num_components, 97 bit_size, srcs); 98 99 nir_load_const_instr *new_instr = 100 nir_load_const_instr_create(mem_ctx, 101 instr->dest.dest.ssa.num_components, 102 instr->dest.dest.ssa.bit_size); 103 104 memcpy(new_instr->value, dest, sizeof(*new_instr->value) * new_instr->def.num_components); 105 106 nir_instr_insert_before(&instr->instr, &new_instr->instr); 107 108 nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, 109 nir_src_for_ssa(&new_instr->def)); 110 111 nir_instr_remove(&instr->instr); 112 ralloc_free(instr); 113 114 return true; 115} 116 117static bool 118constant_fold_intrinsic_instr(nir_intrinsic_instr *instr) 119{ 120 bool progress = false; 121 122 if (instr->intrinsic == nir_intrinsic_discard_if && 123 nir_src_is_const(instr->src[0])) { 124 if (nir_src_as_bool(instr->src[0])) { 125 /* This method of getting a nir_shader * from a nir_instr is 126 * admittedly gross, but given the rarity of hitting this case I think 127 * it's preferable to plumbing an otherwise unused nir_shader * 128 * parameter through four functions to get here. 129 */ 130 nir_cf_node *cf_node = &instr->instr.block->cf_node; 131 nir_function_impl *impl = nir_cf_node_get_function(cf_node); 132 nir_shader *shader = impl->function->shader; 133 134 nir_intrinsic_instr *discard = 135 nir_intrinsic_instr_create(shader, nir_intrinsic_discard); 136 nir_instr_insert_before(&instr->instr, &discard->instr); 137 nir_instr_remove(&instr->instr); 138 progress = true; 139 } else { 140 /* We're not discarding, just delete the instruction */ 141 nir_instr_remove(&instr->instr); 142 progress = true; 143 } 144 } 145 146 return progress; 147} 148 149static bool 150constant_fold_block(nir_block *block, void *mem_ctx) 151{ 152 bool progress = false; 153 154 nir_foreach_instr_safe(instr, block) { 155 switch (instr->type) { 156 case nir_instr_type_alu: 157 progress |= constant_fold_alu_instr(nir_instr_as_alu(instr), mem_ctx); 158 break; 159 case nir_instr_type_intrinsic: 160 progress |= 161 constant_fold_intrinsic_instr(nir_instr_as_intrinsic(instr)); 162 break; 163 default: 164 /* Don't know how to constant fold */ 165 break; 166 } 167 } 168 169 return progress; 170} 171 172static bool 173nir_opt_constant_folding_impl(nir_function_impl *impl) 174{ 175 void *mem_ctx = ralloc_parent(impl); 176 bool progress = false; 177 178 nir_foreach_block(block, impl) { 179 progress |= constant_fold_block(block, mem_ctx); 180 } 181 182 if (progress) { 183 nir_metadata_preserve(impl, nir_metadata_block_index | 184 nir_metadata_dominance); 185 } else { 186#ifndef NDEBUG 187 impl->valid_metadata &= ~nir_metadata_not_properly_reset; 188#endif 189 } 190 191 return progress; 192} 193 194bool 195nir_opt_constant_folding(nir_shader *shader) 196{ 197 bool progress = false; 198 199 nir_foreach_function(function, shader) { 200 if (function->impl) 201 progress |= nir_opt_constant_folding_impl(function->impl); 202 } 203 204 return progress; 205} 206